Confidence-Based Undertrained Network Switching for Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks struggle to accurately classify images significantly different from training sample images, limiting their application in critical areas like full driving automation and fully-autonomous surgery.
Innovation Solution
A neural network architecture that includes an image input device, feature identification unit, feature classification unit, feature classification evaluator, and image classification unit, with a mechanism to switch to an undertrained network if the feature percentages do not exceed a prespecified value, allowing it to correctly classify images with high confidence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network is trained with manually supplied training sample images, then the network can correctly classify images similar to the training samples, but it cannot classify images significantly different from the training samples
Solution Approach 1:
The system dynamically switches between a trained network and an undertrained network based on the classification confidence level. When the trained network's confidence is below a threshold, the system transitions to using an undertrained network, allowing adaptive response to different image types and orientations without retraining the entire network.
Solution Approach 2:
The system changes the network training parameter by using an undertrained network (with fewer training epochs) when the trained network shows low confidence. This parameter change allows the system to explore alternative feature representations that may be more suitable for images significantly different from training samples.
2Quantity of substance
If data augmentation is used to generate variations of training samples, then the number of training samples increases, but the neural network still cannot correctly classify images significantly different from training samples
Solution Approach 1:
Rather than relying solely on static augmented training data, the system dynamically selects between trained and undertrained networks based on real-time confidence assessment. This dynamic selection provides adaptability to image variations that static augmentation cannot capture.
3Measurement precision
If the neural network is fully trained to improve accuracy on training samples, then over-fitting occurs, but classification accuracy on significantly different images remains poor
Solution Approach 1:
The system introduces an intermediary mechanism (the undertrained network) that acts as a mediator when the trained network's confidence is insufficient. This intermediary provides an alternative classification pathway that has not over-fit to training samples, thereby improving reliability for different image orientations.
Solution Approach 2:
The system changes the training parameter by switching to an undertrained network when confidence is low. This parameter change effectively reduces the degree of over-fitting for specific image types, improving generalization to different orientations and variations.
Data Source
AI summary
A neural network for image classification. An image input device receives an image. A feature identification unit identifies a plurality of features, and sends a plurality of identified features. A feature classification unit classifies the identified features, and sends a plurality of feature classifications. A feature classification evaluator sends an image classification initiation signal if sum of the feature percentages associated with one of the feature classifications exceeds a prespecified value, and sends the undertrained network selection signal otherwise. An image classification unit generates an image classification.